Microsoft product manager Keiji Kanazawa allowed autonomous agents to refactor his software for three weeks, only to discover broken features, hollow architectures, and blown token budgets. He rebuilt the system using automated evaluation rubrics rather than manual code reviews.